[#17156][fix] Flush buffered text in DeepSeekR1Parser.finish()

[Yigtwxx]

Description

Fixes #17156.

DeepSeekR1Parser.parse_delta withholds a trailing fragment that could still grow into
a <think> / </think> tag, keeping it in self._buffer until the next delta arrives.
BaseReasoningParser.finish() exists so a parser can flush that state when the stream
ends, and the serving layer calls it (serve/postprocess_handlers.py:177,
serve/responses_utils.py:963). DeepSeekR1Parser never overrode finish(), so a
stream that ended while a fragment was buffered silently dropped those characters from
content or reasoning_content — for example a response ending in a literal <, or
one truncated by max_tokens partway through </thin.

This adds the missing finish() override: the withheld text is emitted and attributed
to the block it was withheld in (reasoning content inside a reasoning block, visible
content otherwise). A buffer holding exactly a complete tag is a delimiter rather than
model output, so it is still discarded, which keeps the existing behavior for a stray
closing tag arriving as the final delta.

This is a conformance fix rather than a behavior change: two sibling parsers in the same
file already implement exactly this flush — NemotronV3ReasoningParser.finish() and
Gemma4ReasoningParser.finish() — and Gemma4ReasoningParser.finish() has the same
shape as the implementation added here. parse_delta is untouched.

Scope: every parser key backed by DeepSeekR1Parser (deepseek-r1, qwen3, qwen3_5,
laguna, minimax_m2, minimax_m2_append_think) plus the subclasses
MiniMaxM3ReasoningParser (minimax_m3) and DeepSeekV4ReasoningParser
(deepseek_v4), whose finish() delegates to the base parser and was a no-op until now.
NemotronV3ReasoningParser overrides finish() in full and is unaffected.

Test Coverage

tests/unittest/llmapi/test_reasoning_parser.py (CPU-only, no model weights):

• test_deepseek_r1_reasoning_parser_finish_flushes_reasoning — a withheld partial
  closing tag is emitted as reasoning content at end of stream. Parametrized over the
  keys that stream from inside the reasoning block: deepseek-r1, qwen3_5,
  minimax_m2, minimax_m2_append_think. qwen3_5 and minimax_m2_append_think were
  not referenced by any test in this file before.
• test_deepseek_r1_reasoning_parser_finish_flushes_content — a withheld partial opening
  tag is emitted as visible content, for qwen3 and laguna.
• test_deepseek_r1_reasoning_parser_finish_drops_complete_tag — a buffer holding only a
  delimiter does not leak into the output.
• test_deepseek_r1_reasoning_parser_finish_is_idempotent — a second finish() returns
  an empty result.
• test_deepseek_r1_reasoning_parser_stream_matches_non_stream — streaming one character
  at a time and then finishing produces the same content / reasoning_content split as
  parse() on the whole text. This is the contract the missing flush violated.

Results on this branch: 173 passed. Against main with only the test changes applied,
19 of the new cases fail, so they do guard the fix. All 136 pre-existing cases in the
file pass unchanged.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

Dev Engineer Review

• DeepSeekR1Parser.finish() flushes buffered text at end-of-stream.
• Complete <think> and </think> tags remain discarded.
• Partial tag fragments emit to the active reasoning_content or content field.
• The implementation applies to parser aliases and delegating subclasses.
• The method is idempotent and preserves streaming/non-streaming equivalence.
• No configuration or test-list files changed.

QA Engineer Review

• Added tests for:
  • Buffered reasoning-content flushing.
  • Buffered visible-content flushing.
  • Parser aliases and delegating subclasses.
  • Complete delimiter suppression.
  • Idempotent finish().
  • Streaming and non-streaming output equivalence.
• No test-list coverage entry was added or modified.
• Verdict: needs follow-up because test-db/ and qa/ coverage data is unavailable.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 2f2a77b7-c013-44ca-831a-a7b6368be7b0

📥 Commits

Reviewing files that changed from the base of the PR and between cfbc7ce and f61b492.

📒 Files selected for processing (1)

• tests/unittest/llmapi/test_reasoning_parser.py

🚧 Files skipped from review as they are similar to previous changes (1)

• tests/unittest/llmapi/test_reasoning_parser.py

---

Walkthrough

DeepSeekR1Parser.finish() now flushes buffered streaming fragments based on parsing state and discards complete delimiters. Tests cover aliases, idempotency, delimiter handling, and parity with non-streaming parsing.

Changes

Reasoning parser end-of-stream handling

Layer / File(s)	Summary
Buffered-text finalization tensorrt_llm/llmapi/reasoning_parser.py	DeepSeekR1Parser.finish() emits incomplete buffered text as reasoning or visible content and discards complete <think> and </think> delimiters.
Finalization behavior tests tests/unittest/llmapi/test_reasoning_parser.py	Tests cover parser aliases, reasoning and content flushing, delimiter suppression, repeated finish() calls, and equivalence between streaming and non-streaming parsing.

Estimated code review effort: 2 (Simple) | ~10 minutes

Suggested reviewers: junyixu-nv, allisonlim-nv

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies issue #17156, the fix type, and the buffered-text flush in DeepSeekR1Parser.finish().
Description check	✅ Passed	The description explains the defect, solution, affected parser keys, tests, and checklist status.
Linked Issues check	✅ Passed	The implementation and tests satisfy issue #17156 by flushing partial buffered fragments while discarding complete delimiter tags.
Out of Scope Changes check	✅ Passed	The changes are limited to the required parser fix and focused tests for the linked issue objectives.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧹 Nitpick comments (2)

tests/unittest/llmapi/test_reasoning_parser.py (2)

78-79: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Add complete type annotations to the new test functions.

Add -> None to each test function. Use list[str] for delta_texts.

As per coding guidelines, "Annotate every function" and "use precise ... types."

Also applies to: 95-96, 108-109, 119-119, 135-135

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/unittest/llmapi/test_reasoning_parser.py` around lines 78 - 79, Update
the new test functions, including
test_deepseek_r1_reasoning_parser_finish_flushes_reasoning and the additional
functions at the referenced locations, with complete annotations: use list[str]
for delta_texts and add -> None to each function signature.

Source: Coding guidelines

---

65-67: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Make the test-only constant private.

R1_AT_START_KEYS only supports parametrization in this module. Rename it to _R1_AT_START_KEYS.

As per coding guidelines, "Prefix non-public names with _."

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/unittest/llmapi/test_reasoning_parser.py` around lines 65 - 67, Rename
the test-only constant R1_AT_START_KEYS to _R1_AT_START_KEYS and update every
reference to it in the module, preserving its parametrization behavior.

Source: Coding guidelines

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@tests/unittest/llmapi/test_reasoning_parser.py`:
- Around line 78-79: Update the new test functions, including
test_deepseek_r1_reasoning_parser_finish_flushes_reasoning and the additional
functions at the referenced locations, with complete annotations: use list[str]
for delta_texts and add -> None to each function signature.
- Around line 65-67: Rename the test-only constant R1_AT_START_KEYS to
_R1_AT_START_KEYS and update every reference to it in the module, preserving its
parametrization behavior.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 88d07722-0ea2-4ab1-9d81-098fc96cceae

📥 Commits

Reviewing files that changed from the base of the PR and between a9544e0 and cfbc7ce.

📒 Files selected for processing (2)

• tensorrt_llm/llmapi/reasoning_parser.py
• tests/unittest/llmapi/test_reasoning_parser.py